ACIH-VQT: aesthetic constraints incorporated hierarchical VQ-transformer for text logo synthesis

Zhixiong Liu, Fang Liu, Mohan Zhang, Shenglan Cui

Abstract: Text logos are essential for brand representation, requiring a balance between aesthetics and readability. However, existing automated methods struggle to meet the unique requirements of text logo synthesis, such as custom font styles and the aesthetic integration of background images. To bridge the gap between typographic characteristics and spatial visual semantics in text logo synthesis, while ensuring both aesthetic appeal and readability, we propose Aesthetic Constraints Incorporated Hierarchical Vector-Quantized Transformer (ACIH-VQT), a novel framework for image content-aware text logo synthesis. The ACIH-VQT leverages a discrete variational autoencoder and hierarchical vector quantization to generate visually appealing and image content-aware text logos. To ensure adherence to design principles, aesthetic constraints are incorporated into encoding training stages. We conduct extensive experiments on the general TextLogo3K dataset. The results demonstrate the effectiveness of ACIH-VQT in generating visually attractive text logos.